Nvidia backs Marvell expansion
Nvidia has invested roughly $2 billion in Marvell to pursue custom XPUs, high‑speed interconnects and silicon‑photonics work, extending Nvidia’s reach beyond accelerators into networking and optical links. The move signals a push to control more of the compute stack that connects AI accelerators at scale. (fool.com)
Nvidia said on March 31 that it invested $2 billion in Marvell and expanded their partnership around custom artificial intelligence chips, networking and optical links. (investor.nvidia.com) The companies said Marvell will supply custom XPUs — processors built for specific workloads — plus networking that works with Nvidia’s NVLink Fusion system. Nvidia said it will contribute its Vera central processing unit, ConnectX network interface cards, BlueField data processing units, NVLink interconnect and Spectrum-X switches. (nvidianews.nvidia.com) They also said they will work on silicon photonics, which sends data with light instead of copper traces, and on optical interconnects for artificial intelligence systems and telecom networks. The same March 31 announcement tied the partnership to Nvidia Aerial artificial intelligence radio access network work for fifth-generation and sixth-generation wireless networks. (investor.marvell.com) The immediate problem is not only building more accelerators, but moving data between them fast enough inside a rack and across a data center. Nvidia’s pitch is that NVLink Fusion lets customers build semi-custom systems that still plug into Nvidia’s broader stack. (investor.nvidia.com) That gives Marvell a larger role in the part of the artificial intelligence market that sits next to the graphics processing unit: custom silicon, switching, digital signal processing and optics. Marvell told investors on March 5 that fiscal 2026 revenue reached a record $8.195 billion, up 42% year over year, driven by artificial intelligence demand. (investor.marvell.com) Marvell has been preparing investors for that shift for months. On its custom artificial intelligence investor event page, the company said it was showcasing a platform for “the next generation of custom AI infrastructure” and the “growing opportunity for custom silicon.” (marvell.com) Nvidia, for its part, is opening a system it once kept centered on its own chips. In the March 31 release, the company said NVLink Fusion is meant to give customers “greater choice and flexibility” while keeping those systems fully compatible with Nvidia infrastructure. (nvidianews.nvidia.com) Jensen Huang said on March 31 that “token generation demand is surging,” and Marvell Chief Executive Matt Murphy said the expanded partnership reflects the importance of “high-speed connectivity, optical interconnect and accelerated infrastructure” in scaling artificial intelligence. The deal leaves Nvidia deeper inside the hardware that connects artificial intelligence chips, not just the chips themselves. (investor.nvidia.com)
